How an idea earns the right to be discussed
Every project begins with a proposed market mechanism, a falsifiable claim, and a statement of what evidence would change the conclusion. The objective is not to rescue an attractive backtest. It is to determine whether a potential return source remains credible after the convenient assumptions are removed.
1. Define the edge before measuring it
The research states why the behaviour could exist, who may be paying for it, why it may persist, and which simpler exposure could explain it. The data universe, timing convention, signal definition, portfolio rule, and rejection criteria are made explicit.
2. Separate discovery from validation
Model selection and parameter discovery belong in-sample. Assessment belongs in held-out or out-of-sample periods, with walk-forward analysis where the use case warrants it. Comparisons include simple baselines so complexity must earn its place.
3. Test fragility, not one preferred specification
Validation examines parameter neighbourhoods, subperiods, instrument subsets, missing data, alternative definitions, and plausible regime partitions. Multiple testing and selection bias are considered when many ideas or variants were searched. A narrow optimum is treated as a warning, not an achievement.
4. Make market frictions part of the result
Turnover, fees, spreads, slippage, latency, borrow, liquidity, rebalancing, and market impact are included when relevant. Capacity and operational requirements are discussed before a statistically attractive result is described as institutionally useful.
5. Evaluate robustness across regimes
The work inspects behaviour across volatility, trend, liquidity, rate, and stress environments when the available history supports those distinctions. No finite sample proves permanence; the purpose is to identify dependence and likely failure states.
6. Judge the idea inside a portfolio
Portfolio relevance includes correlation stability, marginal risk contribution, drawdown overlap, tail dependence, sizing sensitivity, and interaction with exposures already owned. Diversification is a measured behaviour, not a marketing label.
Human review and AI assistance
Automated research systems may help collect sources, structure analysis, run reproducible checks, or prepare a draft. AI-assisted work remains private until a human publisher checks attribution, methodology, evidence, citations, disclosure, and limitations.
7. Publish the failure conditions
No process eliminates model risk, data error, regime change, crowding, or implementation uncertainty. Every dossier should state what can break, what remains untested, and what evidence would require the thesis to be revised or rejected.